Autoware Diffusion Planner

Overview

The Autoware Diffusion Planner is a trajectory generation module for autonomous vehicles, designed to work within the Autoware ecosystem. It leverages the Diffusion Planner model, as described in the paper “Diffusion-Based Planning for Autonomous Driving with Flexible Guidance” by Zheng et al.

This planner generates smooth, feasible, and safe trajectories by considering:

• Dynamic and static obstacles
• Vehicle kinematics
• User-defined constraints
• Lanelet2 map context
• Traffic signals and speed limits

It is implemented as a ROS 2 component node, making it easy to integrate into Autoware-based stacks. The node is aimed at working within the proposed Autoware new planning framework.

Features

• Diffusion-based trajectory generation for flexible and robust planning

  

• Integration with Lanelet2 maps for lane-level context

  

• Dynamic and static obstacle handling using perception inputs

  

  

• Traffic signal and speed limit awareness

  

• ONNX Runtime inference for fast neural network execution
• ROS 2 publishers for planned trajectories, predicted objects, and debug markers

Parameters

{{ json_to_markdown(“planning/autoware_diffusion_planner/schema/diffusion_planner.schema.json”) }}

Parameters can be set via YAML (see config/diffusion_planner.param.yaml).

Inputs

Outputs

Testing

Unit tests are provided and can be run with:

colcon test --packages-select autoware_diffusion_planner
colcon test-result --all

❗ Limitations

While the Diffusion Planner shows promising capabilities, there are several limitations to be aware of:

• Route Termination: The route input to the model consists of a sequence of preferred lanelets from the current position to the goal region. However, this route does not necessarily end exactly at the goal position. As a result, the ego vehicle may continue driving past the goal instead of stopping at the target location.

• Training Dataset Domain Gap: The provided diffusion model checkpoint was trained on datasets using a proprietary Lanelet2 map that is not publicly available. Consequently, performance may significantly degrade when running on other maps, especially in environments with different topology or tagging conventions.

• Route Adherence & Lane Changing: The model sometimes fails to strictly follow the preferred lanelet route. If the ego vehicle leaves the preferred lane (e.g., to avoid an obstacle), it tends to only return to the route during curves. It seldom performs deliberate lane changes to merge back into the correct route on straight segments.

• Agent and Obstacle Avoidance: Although the planner reacts to other agents and can perform avoidance maneuvers, this behavior is not fully reliable. In some cases, collisions with static or dynamic obstacles may still occur due to ignored agents or insufficient context comprehension.

• Lack of Static Object Context: Static environment context such as traffic cones, guard rails, or construction barriers is not currently provided to the model. Instead, an empty tensor is passed in their place, which can lead to limited understanding of occlusions or drivable boundaries.

Development & Contribution

• Follow the Autoware coding guidelines.
• Contributions, bug reports, and feature requests are welcome via GitHub issues and pull requests.

References

• Diffusion Planner (original repo)
• Diffusion planner (our fork of the previous repo, used to train the model)
• “Diffusion-Based Planning for Autonomous Driving with Flexible Guidance”

License

This package is released under the Apache 2.0 License.